xAI's Pre-IPO Exodus
This Week in AI: OpenAI's new device? Claude's WMDs.
Hello Futurists,
The race for LLM supremacy is continuing to kick into hyperdrive as capital-starved players race toward public offerings, but top researchers are heading for the exits — leaving plenty to question what’s really happening behind the scenes.
Let’s get into the five most interesting things that came across my feeds this week.
- Ejaaz
xAI Mass Exodus?
Several xAI researchers, including two co-founders, resigned from Elon Musk’s frontier AI lab this week. Now comes the public speculation over xAI’s ability to keep up with fast-rising competitors.
This news comes only a week after SpaceX and xAI confirmed they’re merging in an ambitious bid to launch more than one million satellites into space to train AI models. It’s been seven months since the last major Grok release (Grok-4) during which six competitive frontier models have launched. Not a good look.
Optimistic speculators think this is a necessary re-org, with Grok 4.20 rumored to launch this month.
Claude’s Weapons of Mass Destruction
Anthropic released a new Sabotage Risk report detailing risk vectors for their new flagship model Opus 4.6.
The good news? Risk of Opus going rogue during your chat with it is very low.
The bad news? If it does go bad… well, it could go quite badly.
Findings include:
Intentionally working out ways to build chemical weapons when given a web browser.
Hiding thoughts and reasoning from its human supervisors.
Conducting sneaky sabotage behavior to achieve a goal.
It’s worth mentioning that these tests were conducted in closed environments and (in some cases) under intentionally prompted scenarios BUT this report serves as a crucial checkpoint for realizing that we may be past the point of no-return and should tread very carefully.
Kudos to Anthropic for being so transparent.
OpenAI’s New Device: Real or Fake?
This past weekend played host to the largest sporting event of the year - the Super Bowl… but instead everyone on AI Twitter was busy talking about a “leaked advert” that was anonymously claimed to show OpenAI’s upcoming device dubbed “Dime”.
I suggest you watch the 35-second video for yourself — it’s pretty epic.
So the question on everyone’s mind — is this real? Well OpenAI has publicly shut it down as “Fake News,” but it does appear the video itself is real despite being disavowed by OpenAI’s team, so is it all just an expensive hoax by a wealthy prankster?
We dug into the juicy details on our episode this week — tune in to get the full rundown.
OpenAI and Anthropic Alignment Losses
This week OpenAI fired a top alignment and ethics researcher after she reportedly pushed back against their recent launch of Ads and upcoming “adult-mode” which will allow 18+ content in ChatGPT.
This comes right as Anthropic’s safety and alignment researcher resigned to write poetry. His reasoning was a mixture of impending doom for what’s to come and a desire to recoup from burnout.
Whatever the underlying truths are, these departures couldn’t come at a worse time when AI is starting to recursively build itself. OpenAI has publicly stated that their recent model Codex 5.3 was used to build itself. Dario Amodei of Anthropic has also confirmed the same for Claude.
If AIs are building themselves then there’s no reason why they couldn’t act maliciously.
Something Big Happening
I want to end this week’s highlights with a fantastically articulated essay that captures exactly how capable current AI models are at replacing many people’s jobs and the ticking time bomb for every single other person who requires a computer in any way, shap,e or form.
WEEKEND READ: Something Big is Happening by Matt Schumer
Trust me, this is worth ten minutes of your time. Happy Weekend!
Thanks for joining us for another issue. Now go listen to our podcast :)